Part 10 (2/2)

The problem with robots is, they still mostly act like machines. Cynthia Breazeal at MIT sums it up: ”Robots today interact with us either as other objects in the environment, or at best in a manner characteristic of socially impaired people. They generally do not understand or interact with people as people. They are not aware of our goals and intentions.”34 She wants to give her robots theory of mind! She wants her robot to understand her thoughts, needs, and desires. If one is building a robot to help the elderly, she continues, ”Such a robot should be persuasive in ways that are sensitive to the person, such as helping to remind them when to take medication, without being annoying or upsetting. It must understand what the person's changing needs are and the urgency for satisfying them so that it can set appropriate priorities. It needs to understand when the person is distressed or in trouble so that it can get help.”

Kismet, the second-generation Cog, is a sociable robot that was built in the lab of Rodney Brooks, director of the MIT Computer Science and Artificial Intelligence Laboratory, predominantly by Cynthia Breazeal when she was Brooks's graduate student. Part of what makes Kismet a sociable robot is that it has large eyes that look at what it is paying attention to. It is programmed to pay attention to three types of things: moving things, things with saturated color, and things with skin color. It is programmed to look at skin color if it is lonely, and bright colors if it is bored. If it is paying attention to something that moves, it will follow the movement with its eyes. It has a set of programmed internal drives that increase until they release certain behaviors. Thus if its lonely drive is high, it will look around until it finds a person. Then, since that drive is satisfied, another drive will kick in, perhaps boredom, which will increase, and it will start searching for a bright color; this makes it appear to be looking for something specific. It may then find a toy, giving an observer the impression that it was looking specifically for the toy. It also has an auditory system that detects prosody in speech. With this mechanism it has a program that matches certain prosody with specific emotions. Thus it can detect certain emotions such as approval, prohibition, attention getting, and soothing-just like your dog. Incoming perceptions affect Kismet's ”mood” or emotional state, which is a combination of three variables: valence (positive or negative), arousal (how tired or stimulated it is), and novelty. Responding to various motion and prosody cues, Kismet will proceed among different emotional states, which are expressed through its eyes, eyebrows, lips, ears, and the prosody of its voice. Kismet is controlled by the interaction of fifteen different computers running various operating systems-a distributed system with no central control. It does not understand what you say to it, and it speaks only gibberish, though gibberish with the proper prosody for the situation. Because this robot simulates human emotions and reactions, many people relate to it on an emotional level and will speak to it as if it were alive. Here we are back to anthropomorphism.

Rodney Brooks wonders if simulated, hard-coded emotions in a robot are the same as real emotions. He presents the argument that most people and artificial intelligence researchers are willing to say that computers with the right software and the right problem can reason about facts, can make decisions, and can have goals; but although they may say that a computer may act as if, behave as if, seem as if, or simulate that it is afraid, it is hard to find anyone who will say that it is viscerally afraid. Brooks sees the body as a compilation of biomolecules that follow specific, well-defined physical laws. The end result is a machine that acts according to a set of specific rules. He thinks that although our physiology and const.i.tuent materials may be vastly different, we are much like robots. We are not special or unique. He thinks that we overanthropomorphize humans, ”who are after all mere machines.”9 I'm not sure that, by definition, it is possible to overanthropomorphize humans. Perhaps it is better to say we underanthropomorphize machines or undermechanomorphize humans.

Breazeal's group's next attempt at developing TOM in a robot is Leonardo. Leo looks like a puckish cross between a Yorks.h.i.+re terrier and a squirrel that is two and a half feet tall.* He can do everything that Kismet can do and more. They wanted Leo to be able to identify another's emotional state and why that person is experiencing it. They also want him (they refer to Leo as ”he” and ”him,” so I will, too) to know the emotional content of an object to another person. They don't want Leo tramping on the Gucci shoes or throwing out your child's latest painting that looks like trash to anyone but a parent. They also want people to find Leo easy to teach. Instead of your having to read an instruction manual and learn a whole new form of communication when you get your first robot, they want Leo to be able to learn as we do. You'll just say, ”Leo, water the tomatoes on Thursdays” and show him how to do it, and that's it. No small ambitions!

They are banking on the neuroscience theory that humans are sociable, and we learn through using our social skills. So first, in order to be responsive in a social way, Leonardo has to be able to figure out the emotional state of the person with whom he is interacting. They approached designing Leo using evidence from neuroscience that ”the ability to learn by watching others (and in particular the ability to imitate) could be a crucial precursor to the development of appropriate social behavior-and ultimately the ability to reason about the thoughts, intents, beliefs, and desires of others.” This is the first step on the road to TOM. The design was inspired by the work done on newborns' facial imitation and simulation ability by Andrew Metzoff and M. Keith Moore, whom we read about in chapter 5. They needed Leonardo to be able to do the five things that we talked about that a baby could do when it was hours old: Locate and recognize the facial features of a demonstrator.

Find the correspondence between the perceived features and its own.

Identify a desired expression from this correspondence.

Move its features into the desired configuration.

Use the perceived configuration to judge its own success.

So they built an imitation mechanism into Leonardo. Like Kismet, he has visual inputs, but they do more. Leo can recognize facial expressions. Leo has a computational system that allows him to imitate the expression he sees. He also has a built-in emotional system that is matched to facial expression. Once this system imitates a person's expression, it takes on the emotion a.s.sociated with it.

The visual system also recognizes pointing gestures and uses spatial reasoning to a.s.sociate the gesture with the object that is indicated. Leonardo also tracks the head pose of another. Together these two abilities allow him to understand the object of attention and share it. He makes and keeps eye contact.

Like Kismet, he has an auditory system, and he can recognize prosody, pitch, and the energy of vocalization to a.s.sign a positive or negative emotional value. And he will react emotionally to what he hears. But unlike Kismet, Leo can recognize some words. His verbal tracking system matches words to their emotional appraisal. For instance the word friend has a positive appraisal, and the word bad has a negative one, and he will respond with the emotional expression that matches the words.

Breazeal's group also incorporated the neuroscience findings that memory is enhanced by body posture and affect.36 As Leo stores information in long-term memory, the memory can be linked with affect. His ability to share attention also allows him to a.s.sociate emotional messages of others with things in the world. You smile as you look at the painting your kid did; Leo looks at it too, and he files it away in memory as a good thing-he doesn't toss it with the trash. Shared attention also provides a basis for learning.

So we are reasonably close to a robot that is physically humanlike in appearance and movement, one that can simulate emotions and is sociable. However, you'd better not be doing the rumba with your robot, because it most likely would break your foot if it accidentally trod on it (these puppies are not lightweight). You should also consider its energy requirements (there goes the electric bill). But what about intelligence? Social intelligence is not all my robot will need. It is going to have to outfox gophers, and it is going to have to be pretty dang intelligent to outfox the gophers in my yard, which, I am sure, have the same genetic code as the Caddyshack survivors.

Ray Kurzweil is not worried so much about the physical vehicle. It is the intelligence that interests him. He thinks that once computers are smart enough, that is, smarter than we are, they will be able to design their own vehicles. Others think that humanlike intelligence and all that contributes to it cannot exist without a human body: I think therefore my brain and my body am. Alun Anderson, editor in chief of New Scientist magazine, put it this way when asked what his most dangerous idea was: ”Brains cannot become minds without bodies.”37 No brain-in-a-box will ever have humanlike intelligence. We have seen how emotion and simulation affect our thinking, and, without those inputs, we would be, well, a whole 'nother animal. And Jeff Hawkins, creator of the Palm Pilot, thinks since we don't even know what intelligence is and what processes in the brain produce it, we have a lot of work still to do before we can have intelligent machines.38 ARTIFICIAL INTELLIGENCE.

The term artificial intelligence (AI) originated in 1956, when John McCarthy from Dartmouth College, Marvin Minsky from Harvard University, Nathaniel Rochester of the IBM Corporation, and Claude Shannon from the Bell Telephone Laboratories proposed that ”a 2 month, 10 man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hamps.h.i.+re. The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.”39 Looking back at that statement made over half a century ago, it seems as if it was a little optimistic. Today the American a.s.sociation for Artificial Intelligence defines AI as ”the scientific understanding of the mechanisms underlying thought and intelligent behavior and their embodiment in machines.”40 However, despite all the computing power and effort that have gone into making computers intelligent, they still can't do what a three-year-old child can do: They can't tell a cat from a dog. They can't do what any surviving husband can do: They don't understand the nuances of language. For instance, they don't know that the question ”Have the trash barrels been taken out?” actually means, ”Take the trash barrels out,” and that it also has a hidden implication: ”If you don't take the trash out, then....” Use any search engine, and as you gaze at what pops up, you think, ”Where did that come from? That is so not what I'm looking for.” Language translation programs are wacky. It is obvious the program has no clue as to the meaning of the words it is translating. Attempts are continually being made, but even with all the processing power, memory, and miniaturization, creating a machine with human intelligence is still a dream. Why?

Artificial intelligence comes in two strengths: weak and strong. Weak AI is what we are used to when we think about computers. It refers to the use of software for problem-solving or reasoning tasks. Weak AI does not include the full range of human cognitive abilities, but it may also have abilities that humans do not have. Weak AI has slowly permeated our lives. AI programs are directing our cell-phone calls, e-mails, and Web searches. They are used by banks to detect fraudulent transactions, by doctors to help diagnose and treat patients, and by lifeguards to scan beaches to spot swimmers in need of help. AI is responsible for the fact that we never encounter a real person when we make a call to any large organization or even many small ones, and for the voice recognition that allows us to answer vocally rather than press a number. Weak AI beat the world champion chess player, and can actually pick stocks better than most a.n.a.lysts. But Jeff Hawkins points out that Deep Blue, IBM's computer that beat the world chess champion, Garry Kasparov, at chess in 1997, didn't win by being smarter than a human. It won because it was millions of times faster than a human: It could evaluate two hundred million positions per second. ”Deep Blue had no sense of the history of the game, and didn't know anything about its opponent. It played chess yet didn't understand chess, in the same way that a calculator performs arithmetic but doesn't understand mathematics.”38 Strong AI is what flips many people out. Strong AI is a term coined by John Searle, a philosopher at the University of California, Berkeley. The definition presupposes, although he does not, that it is possible for machines to comprehend and to become self-aware. ”According to strong AI, the computer is not merely a tool in the study of the mind; rather, the appropriately programmed computer really is a mind, in the sense that computers given the right programs can be literally said to understand and have other cognitive states.”41 Searle maintains that all conscious states are caused by lower level brain processes,42 thus consciousness is an emergent phenomenon, a physical property-the sum of the input from the entire body. Consciousness does not just arise from banter back and forth in the brain. Consciousness is not the result of computation. You have to have a body, and the physiology of the body and its input, to create a mind that thinks and has the intelligence of the human mind.

IS A CONSCIOUS MACHINE POSSIBLE?.

The logic behind believing a machine can be conscious is the same logic that is behind creating AI. Because human thought processes are the result of electrical activity, if you can simulate that same electrical activity in a machine, then the result will be a machine with humanlike intelligence and consciousness. And just as with AI, there are some who think that this does not mean that the machine's thought processes need necessarily be the same as a human's to produce consciousness. Then there are those who agree with Hawkins and think that it must have the same processes, and that to have those, it has to be hooked up the same way. And there are those who are on the fence.

The quest for artificial intelligence was not originally based on reverse-engineering the brain, because in 1956, when AI was a glimmer of an idea, very little was known about how the brain works. Those early engineers had to wing it when they began to design AI. They initially came up with their own solutions for creating the various components of artificial intelligence, and some of these methods have actually supplied clues to how parts of the brain work. Some of these approaches are based on mathematical rules, such as Bayesian logic, which determines the likeliness of a future event based on similar events in the past, or Markov models, which evaluate the chance that a specific sequence of events will happen and are used in some voice-recognition software. The engineers built ”neural nets,” set up to run in parallel and loosely simulating neurons and their connections; they actually learn responses that are not preprogrammed in. These systems have also been used in voice-recognition software. They are also used to detect fraud in credit-card charges, and in face and handwriting recognition. Some are based on inference-the old ”if this, then that” logic. There are programs that search through large numbers of possibilities, such as the chess program Deep Blue. Some are planning programs that start with general facts about the world, rules about cause and effect, facts germane to particular situations, and the intended goal-just like the direction finder in your car that plans routes and tells you how to get to the closest Chinese takeout.

But the human brain is different in many ways from a computer. In his book The Singularity Is Near, Kurzweil enumerates the differences.

The brain's circuits are slower but more ma.s.sively parallel. The brain has about one hundred trillion interneuronal connections. This is more than any computer yet has.

The brain is constantly rewiring itself and self-organizing.

The brain uses emergent properties, which means that intelligent behavior is rather an unpredictable result of chaos and complexity.

The brain is only as good as it has to be, in terms of evolution. There's no need to be ten times smarter than everyone else; you need only be a little smarter.

The brain is democratic. We contradict ourselves: We have internal conflicts that may result in a superior solution.

The brain uses evolution. The developing brain of a baby six to eight months old forms many random synapses. The patterns of connections that best make sense of the world are the ones that survive. Certain patterns of brain connections are crucial, whereas some are random. As a result, an adult has far fewer synapses than the toddler.

The brain is a distributed network. There is no dictator or central processor calling the shots. It is also deeply connected: Information has many ways to navigate through the network.

The brain has architectural regions that perform specific functions and have specific patterns of connections.

The overall design of the brain is simpler than the design of a neuron.2 It's interesting, however, that Kurzweil leaves out something rather major. He ignores the fact that the brain is hooked up to a biological body. So far, AI programs are good only at the thing they are specifically designed for. They don't generalize and aren't flexible.2 Deep Blue, with all its connections, ma.s.sive memory, and power, does not know that it better take the trash out...or else.

Although human-level intelligence has not been achieved, computers surpa.s.s some of our abilities. They are better at symbolic algebra and calculus, scheduling complex tasks or sequences of events, laying out circuits for fabrication, and many other mathematically involved processes.9 They are not good at that elusive quality, common sense. They can't critique a play. As I said before, they are not good at translating from one language to another, nor at nuances within a language. Oddly, it is many of the things that a four-year-old can do, rather than what a physicist or a mathematician can do, that are the hang-ups.

No computer yet has pa.s.sed the Turing Test, proposed in 1950 by Alan Turing,43 the father of computer science, to answer the question, Can machines think? In the Turing Test, a human judge engages in a natural language conversation with two other parties, one a human and the other a machine, both trying to appear human. If the judge cannot reliably tell which is which, then the machine has pa.s.sed the test. The conversation is usually limited to written text, so that voice is not a prejudicial factor. Many researchers have a problem with the Turing Test. They do not think that it will indicate whether a machine is intelligent. Behavior isn't a test of intelligence. A computer may be able to act as if it were intelligent, but that does not mean it is.

PALM PILOT TO THE RESCUE.

Jeff Hawkins thinks he knows why no truly intelligent machines have been made. It is not because computers just need to be more powerful and have more memory, as some researchers think. He thinks everyone working on artificial intelligence has been barking up the wrong tree. They have been working under the wrong premise38 and should be paying more attention to how the human brain works. Although John McCarthy and most other AI researchers think that ”AI does not have to confine itself to methods that are biologically observable,”44 Hawkins thinks this notion is what has led AI research astray. And he isn't so happy with neuroscientists, either. Slogging through neuroscience literature to answer the question of just how the brain works, he found that although mounds of research have been done, and tons of data acc.u.mulated, no one yet has put it all together and come up with a theory to explain how humans think. He was tired of the failed attempts at AI and concluded that if we don't know how humans think, then we can't create a machine that can think like a human. He also concluded that if no one else was going to come up with a theory, he'd just have to do it himself. So he founded the Redwood Center for Theoretical Neuroscience and set about the business. Jeff is no slouch. Or maybe he is. He leaned back, put his feet up on the desk, cogitated, and came up with the memory-prediction theory,38 which presents a large-scale framework of the processes in the human brain. He hopes other computer scientists will take it out for a spin, tweak it, and see if it works.

Hawkins was fascinated when he read a paper written in 1978 by the distinguished neuroscientist Vernon Mountcastle, who had made the observation that the neocortex is remarkably similar throughout, and therefore all regions of the cortex must be performing the same job. Why the end result of that job is different for different areas-that is, vision is the result of processing in the visual cortex, hearing in the auditory cortex, etc.-is not because they have different processing methods. It is because the input signals are different, and because of how the different regions are connected to each other.

One piece of evidence that backs up this conclusion was the demonstration of the plasticity (an ability to change its wiring) of the cortex done by Mriganka Sur at MIT. To see what effect the input to a cortical area had on its structure and function, he rewired visual input in newborn ferrets so that it went to the auditory cortex instead of the visual cortex.45, 46 Would a ferret be able to use another portion of the somatosensory cortex, such as the auditory cortex tissue, to see? It turns out that the input has a big effect. The ferrets could see to some extent. This means that they were seeing with the brain area that normally hears sounds. The new ”visual cortical tissue” isn't wired exactly as it would have been in the normal visual cortex, leading Sur and his colleagues to conclude that input activity can remodel the cortical networks, but it is not the only determinant of cortical structure; there are probably intrinsic cues (genetically determined) that also provide a scaffold of connectivity.47 That means specific areas of the cortex have evolved to process certain types of information and have been wired in a certain way to better accommodate it, but if need be, because the actual mode of processing is the same in all the neurons, any part of the cortex can process it.

This idea that the brain uses the same mechanism to process all information made a lot of sense to Hawkins. It united all the capabilities of the brain into one tidy package. The brain didn't have to reinvent the wheel every time it expanded its abilities: It has one solution for thousands of problems. If the brain uses a single processing method, then a computer could too, if he could figure out what that method was.

Hawkins is a self-declared neocortical chauvinist. He looks on the neocortex as the seat of our intelligence: It was the last to develop and is larger and better connected than any other mammal's. However, he fully keeps in mind that all the input that goes into it has been processed by lower-level brain regions: those regions that are evolutionarily older, which we share with other animals. So using his big neocortex, Hawkins came up with his memory-prediction theory, and we are going to check it out.

All the inputs into the neocortex come from our senses, just as in all animals. One surprising thing is that no matter what sense we are talking about, the input into the brain is in the same format: neural signals that are partly electrical and partly chemical. It is the pattern of these signals that determines what sense you experience; it doesn't matter where they come from. This can be ill.u.s.trated by the phenomenon of sensory subst.i.tution.

Paul Bach y Rita, who was a physician and neuroscientist at the University of Wisconsin, became interested in the plasticity of the brain after caring for his father, who was recovering from a stroke. He understood that the brain is plastic and that it is the brain that sees, not the eyes. He wondered if he could restore vision to a blind person by providing the correct electrical signal but through a different input pathway, that is, not through the eyes, which were

<script>